Chinese AI labs DeepSeek and Moonshot AI launch new frontier models at low prices

Two Chinese AI labs have released new frontier models in quick succession that are shifting the balance of power in the global large language model market. DeepSeek launched model V4-Flash-0731 in public beta via its API on 31 July 2026, while Moonshot AI introduced its Kimi K3 model. Both models combine performance comparable to the top of the American offering with API rates that are considerably lower.

For developers and companies building with AI, the price differences are tangibly felt. DeepSeek V4 is reported to significantly undercut the rates of Anthropic's comparable models, directly affecting decisions around model integration and infrastructure costs. The releases follow a series of Chinese model launches that have drawn attention since early 2025, starting with DeepSeek-R1 in January of that year.

Details on the exact benchmark scores of Kimi K3 and the full price comparison had not been fully verified at the time of writing. The available information on DeepSeek V4-Flash-0731 is more concretely documented and forms the focus below.

DeepSeek V4-Flash-0731: architecture and intended use

DeepSeek V4-Flash-0731 is built as a Mixture-of-Experts (MoE) model with a total of 284 billion parameters. Per token, 13 billion parameters are actively used, which limits the computational load per inference step compared to models with a fully dense architecture. The context window is 1 million tokens, ample for extensive codebases or long document sequences.

According to DeepSeek, the model is specifically aimed at coding, tool use, and so-called agentic workflows, in which an AI system executes tasks across multiple consecutive steps without human intervention. Enhanced agentic capabilities are explicitly cited as a design priority. This aligns with the broader trend of frontier models being deployed not merely as chatbots, but as components of automated workflows.

DeepSeek was founded in July 2023 by Liang Wenfeng, who also serves as CEO. The company, formally named Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., is headquartered in Hangzhou and is funded by High-Flyer, a Chinese hedge fund also led by Liang.

Pricing as a strategic instrument

What sets the Chinese releases apart from many of their American counterparts is the approach to pricing. DeepSeek V4 models are offered at rates that, according to available comparisons, are substantially lower than those of Anthropic's Claude models at a comparable positioning. Exact cents-per-token figures vary by source and had not been uniformly confirmed at the time of publication, but the direction is consistent: API costs are several multiples lower.

Aggressive pricing is not a new pattern for DeepSeek. With the introduction of DeepSeek-R1 in January 2025, the company already set a similar tone, which at the time prompted revisions among Western providers. A structurally lower cost base, partly stemming from the MoE architecture and efficient training methods, makes it possible to gain market share without immediate short-term profit pressure.

For teams deploying models at scale, such as in SaaS products or at high inference volumes, these price differences quickly translate into substantial monthly cost savings. This makes it increasingly difficult for many builders to overlook Chinese providers, even when questions exist around data privacy or export controls.

Moonshot AI Kimi K3: limited details available

Less public information is currently available about Kimi K3, the new model from Moonshot AI, than about DeepSeek V4-Flash-0731. Moonshot AI is a Chinese AI company that previously gained recognition with its Kimi assistant, focused on long-context processing. Kimi K3 is positioned as a frontier model performing at the level of the international top, but here too, fully verified benchmark results from independent evaluations are currently lacking.

The simultaneous launch of multiple Chinese frontier models does reinforce the broader pattern: Chinese labs are releasing models at an ever-faster pace that match what American labs offer at the GPT-4 class or Claude Sonnet class level in terms of architecture and capabilities, and are doing so at a price point that puts Western competition under pressure. Whether Kimi K3 also holds up in that comparison in independent tests will become clearer in the coming weeks.

What this means for those building with or investing in AI

The rise of Chinese frontier models presents developers, product teams, and investors with a trade-off that goes beyond technical performance alone. On one hand, models such as DeepSeek V4-Flash-0731 offer attractive costs and serious capabilities for coding and agentic applications. On the other hand, considerations around data sovereignty, export regulations, and dependence on Chinese infrastructure play a role, particularly for European companies working with sensitive data or operating in regulated sectors.

For investors in the AI ecosystem, the releases demonstrate that competition at the model layer is moving more broadly and rapidly than was anticipated two years ago. The notion that a handful of American labs has a structural hold on the frontier market is becoming increasingly difficult to sustain. At the same time, transparency about actual training costs, energy consumption, and hardware used is something Chinese labs, like their American counterparts, are selective about.

In the coming weeks, independent evaluations of both DeepSeek V4-Flash-0731 and Kimi K3 will provide more clarity on where these models actually stand in practice relative to the established names.